gorbatjovy/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic
gorbatjovy/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic is a quantized version of the Qwen3.6-27B-Architect-Polaris2-Fable-B-F451 model, developed by gorbatjovy. This model is notable for its low refusal rate of 4/100 and a KL divergence of 0.0074, indicating a high degree of fidelity to the original model while potentially offering improved usability. It is suitable for applications requiring a robust and less restrictive language model.
Loading preview...
Model Overview
This model, gorbatjovy/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451-heretic, is a quantized version of the nightmedia/Qwen3.6-27B-Architect-Polaris2-Fable-B-F451 base model. It has been specifically optimized to address certain operational aspects, including a notable fix to its MTP (Multi-Turn Prediction) layers, thanks to contributions from @DavidAU.
Key Characteristics
- Low Refusal Rate: Achieves a refusal rate of 4 out of 100, suggesting a highly compliant and less restrictive response generation compared to many other models.
- Low KL Divergence: Features a Kullback-Leibler (KL) divergence of 0.0074, indicating that the quantized version maintains a very close statistical distribution to its original, unquantized counterpart.
- MTP Layer Fixes: Incorporates critical fixes for its MTP layers, which can enhance the model's performance in multi-turn conversations and complex reasoning tasks.
Use Cases
This model is well-suited for applications where a balance between performance and resource efficiency is desired, particularly in scenarios requiring a language model with a low tendency to refuse prompts. Its low KL divergence ensures that the benefits of quantization do not significantly compromise the original model's capabilities. The MTP layer improvements make it potentially more robust for interactive and conversational AI systems.